biCourage: ngram and syntax GCNs for Hate Speech detection

Hate Speech identification is a challenging task given the world knowledge required. Moreover, it is even more complex in the social media context due to language and media specificities. Despite these challenges, advances in this task may help improving collective well-being on social media. In this context, the biCourage team participated in the English version of Task 1 of HASOC 2021, a shared task for “Hate Speech and Offensive Content Identification in English and Indo-Aryan Languages”. Our participation in this campaign aimed to examine the suitability of Graph Convolutional Neural Networks (GCN), due to their capability to integrate flexible contextual priors, as a computationally effective solution compared to more computationally expensive and relatively data-hungry methods, such as fine-tuning. Specifically, we explored and combined two text-to-graph strategies based on different language modelling objectives, comparing them with fine-tuned Bert. We submitted the results of several deep learning architectures, comprised of different arrangements of GCNs and transformer architectures. Our team achieved the best results in both subtasks using the GCNs based architectures combining two text-to-graph strategies ranked in 21st and 20th positions in Subtasks 1A and 1B. Assessing the models’ prediction, we identify complementary capabilities in the text-to-graph strategies that further research on their combination can explore. Moreover, the proposed GCN model is 3.85 times faster than fine-tuned Bert in training speed and still outperforms it by 2.3% and 5.41% on the F1 score of Subtasks 1A and 1B, respectively.